Recursive Self-Improvement and the Future of AI
发布时间 来源
Episode 设置
本播客单集发布于2026年10-12，邀请了嘉宾 owen 和 grace，共同探讨人工智能中递归自我改进（RSI）的概念。
对话开始时，owen 和 grace 介绍了他们的背景。owen 学习数学，后来从事定理证明机器、持续学习和元学习的研究，之后在 DeepMind 担任了几年研究科学家。grace 学习物理学，并在麻省理工学院媒体实验室获得了强化学习博士学位。
主持人与嘉宾探讨了 RSI 的定义，他们指出这是一个含义过多的术语。他们将真正的 RSI 与单纯的自催化区分开来——在自催化中，人类使用人工智能工具来开发更好的人工智能，但想法和改进仍然源自人类大脑。相比之下，真正的 RSI 涉及将优化器与被优化对象进行对调，其中语言模型和智能体的大脑充当被优化的对象。这形成了一个闭环，使认知改进能够重新投入自身，从而改变能力提升的速度。
嘉宾认为，RSI 不仅是一种理论信念，而且是一个经验性问题。他们的方法依赖于一个多智能体系统，其中人类研究人员和工程师被视为智能体的一个特殊案例，协同工作并进行递归改进。他们不仅关注性能（分子），还强调资本和资源效率（分母）的重要性。他们保持着严格的内部文化，包括一个用于追踪实验成本的“出血板”（bleeder board），以及一名被昵称为“方差总督”（viceroy of variance）的团队成员，其工作是挑战和测试结果，以消除噪声和选择性筛选。
目前，该公司正在将 RSI 应用于推理问题，旨在将其打造成一个高实用性且在经济上重要的领域，从而同时改善分子和分母。虽然他们承认物理限制——例如光速、宇宙中的粒子数量以及计算的最终极限——但他们报告说，在过去的几个月里，他们看到了复合改进的真实、经验性的微光，即在需要更少能量的同时结果变得更好。他们强调，智能是一个多维向量，虽然像编码和数学这样评估快速且廉价的领域可以快速加速，但需要湿实验室实验的物理领域仍将受到现实世界的瓶颈制约。
This podcast episode, published on 2026-10-12, features a discussion with guests owen and grace about the concept of recursive self-improvement (RSI) in artificial intelligence.
The conversation begins with owen and grace introducing their backgrounds. owen studied math and later worked on theorem proving machines, continued learning, and meta-learning before spending a few years at DeepMind as a research scientist. grace studied physics and earned a PhD in reinforcement learning at the MIT Media Lab.
The hosts and guests tackle the definition of RSI, which they note is an overloaded term. They distinguish true RSI from mere autocatalysis—where humans use AI tools to develop better AI, but the ideas and improvements still originate from human brains. In contrast, true RSI involves swapping the optimizer with the optimizee, where the language model and the agent's brain serve as the thing being optimized. This creates a closed loop that allows cognitive improvement to be reinvested back into itself, shifting the rate of capability improvement.
The guests argue that RSI is not just a theoretical belief but an empirical problem. Their approach relies on a multi-agent system where human researchers and engineers are treated as a special case of agents, working together and recursively improving. Rather than just focusing on performance (the numerator), they emphasize the importance of capital and resource efficiency (the denominator). They maintain a strict internal culture, including a "bleeder board" to track the cost of experiments and a team member nicknamed the "viceroy of variance" whose job is to challenge and test results to eliminate noise and cherry-picking.
Currently, the company is applying RSI to the inference problem, aiming to make it a high-utility and economically important area that improves both the numerator and the denominator. While they acknowledge physical limits—such as the speed of light, the number of particles in the universe, and the ultimate bounds of computation—they report seeing real, empirical glimmers of compounding improvements over recent months, where results get better while requiring less energy. They stress that intelligence is a multi-dimensional vector, and while domains with fast and cheap evaluations like coding and mathematics can accelerate rapidly, physical domains requiring wet lab experiments will remain bottlenecked by the real world.
摘要
a16z General Partner Martin Casado sits down with MakerMaker.AI co-founders Dhaval Adjodah and Owen He to unpack recursive self-improvement, or RSI, and separate the increasingly popular term from what it would actually mean for an AI system to improve itself.
Dhaval and Owen distinguish true RSI from simply using AI to automate pieces of AI development. Their goal is a closed loop where the system being improved is also doing the improving, across components including models, harnesses, and eventually more of the AI stack. They explain why closing that loop still doesn’t guarantee runaway “takeoff,” and why measuring real progress requires accounting for both capability gains and the resources required to achieve them.
They also share what they’re seeing in their own research, including early signs that self-improving systems can increase performance while improving efficiency. And they dig into the practical challenges of proving those gains are real, from variance and reward hacking to the enormous difficulty of distinguishing genuine improvement from noise.
Timestamps:
00:00 - Intro
00:58 - What actually counts as RSI
05:10 - Capability as a numerator and denominator
06:58 - Closed loop vs takeoff
08:32 - Humans as agents in the system
13:39 - Why most RSI papers don't scale
17:41 - Why they chose inference
24:00 - "It's not a belief, it's happening"
25:42 - The Viceroy of Variance
29:39 - The garden hose company in Wisconsin
Resources:
Follow Dhaval Adjodah on LinkedIn: https://www.linkedin.com/in/rand0m
Follow Owen He on LinkedIn:https://www.linkedin.com/in/owen-he-b7b064279/
Follow Martin Casado on X: https://x.com/martin_casado
Stay Updated:
If you enjoyed this episode, be sure to like, subscribe, and share with your friends!
Find a16z on X: https://twitter.com/a16z
Find a16z on LinkedIn: https://www.linkedin.com/company/a16z
Listen to the a16z Show on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX
Listen to the a16z Show on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711
Follow our host: https://x.com/eriktorenberg
Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures.
GPT-4正在为你翻译摘要中......
